SC2 AI Bots: Infinite Tournament and Micro Hypothesis Lab

SC2 AI Bots is a research-and-engineering project around autonomous StarCraft II agents. It combines actual bot development, a continuously running AI Arena-style tournament, tactical micro-management experiments, immutable release packages, replay analysis, structured logs, and public dashboards.

NexusBot Development

The core bot is NexusBot, an experimental Protoss agent built on Ares SC2 and python-sc2. The bot evolved from deterministic build-order execution into a modular decision system with economy control, scouting memory, rush detection, adaptive defence, expansion management, tactical squads, target selection, Blink/Stalker micro, Adept shades, cleanup states, recovery logic, and structured decision journals.

Infinite Tournament

The project runs a permanent headless benchmark arena with no seasons, finals, or pauses. Each registered immutable bot package enters calibration matches, then joins a weighted random Elo pool against historical NexusBot releases, test fixtures, and external/control bots. The tournament records W/D/L, Elo, failures, bot crashes, current matches, recent results, replays, and technical diagnostics without treating infrastructure failures as game losses.

Micro Hypothesis Lab

A separate Battle Lab isolates tactical combat from economy. Fixed army compositions fight paired baseline-vs-managed scenarios with identical map, seed, side, and formation. This makes it possible to test hypotheses such as defensive Stalker Blink, Colossus screen positioning, Immortal armored-target preference, Guardian Shield timing, Phoenix lift reservations, Void Ray target locks, firepower-based target scoring, and squad-level focus policies.

Experiment Discipline

The system treats a single attractive replay as anecdotal, not proof. Changes move through unit tests, fixture matches, reproducible smoke runs, repeated seeds, previous-version comparisons, replay inspection, structured logs, and tournament registration. Benchmarks document commit hashes, package names, maps, player order, seeds, repeats, raw results, aggregated metrics, and observed regressions.

Observability

The public dashboard shows runner health, active match state, active bots, rated games, Elo standings, match feed, newcomers, technical failures, and auto-refresh status. Battle Lab exposes micro uplift, managed-vs-baseline rates, remaining army value, damage deltas, duration deltas, matchup matrices, army compositions, recent paired results, and ability telemetry. A browser replay player renders prepared SC2 matches as a pixel-art tactical map with playback controls, timeline events, player stats, economy bars, army value and bot decision streams. Prometheus and Grafana monitor throughput, draw rate, failure rate, runner liveness, stuck matches, and tournament progress.

AI Agent Angle

The project is a practical laboratory for agentic systems under real-time constraints. It connects planning, perception, memory, bounded decision rules, adversarial evaluation, regression testing, telemetry, and continuous competition. It also includes an experimental LLM strategist path for bounded economy and army planning, while keeping the real-time game loop non-blocking and safe when network calls fail.

Product Value

Instead of claiming that a bot became smarter after a code change, the platform makes progress measurable. It turns SC2 bot development into an operating system for hypotheses: build a bot, package it immutably, run it forever, compare it against history, inspect the replay, and promote only changes that survive evidence.